Voice AI interaction state management method for low-power-consumption terminal

Abstract

The invention discloses a voice AI interaction state management method for a low-power-consumption terminal, and aims to solve the problems of unbalanced energy consumption and response speed, lack of interaction content and high hardware cost of voice AI interaction of the low-power-consumption terminal in the prior art. According to the method, a'dormancy-standby-wakeup-interaction 'multi-stage state machine closed-loop management system is constructed, and a dynamic power consumption adjustment algorithm and a'local wakeup-cloud reasoning' mixed architecture are combined. Rapid triggering and intention confirmation are realized locally through hardware-level wake-up word detection and voice activity detection, an LTE module is activated only after effective wake-up to establish cloud connection, and a cloud pre-trained language large model supports flexible semantic analysis and interaction content generation. Meanwhile, a module on-demand start-stop strategy is adopted, power supply of a redundant module is cut off in an inactive stage, a communication link maintains connection with a low-volume heartbeat packet, and the communication link rapidly falls back to a low-power-consumption state after interaction. According to the invention, collaborative optimization of low power consumption and AI interactive experience is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of voice interaction of intelligent terminal equipment, and particularly relates to a voice AI interaction state management method for low-power terminals. BACKGROUND

[0002] The current voice AI interaction system is mainly developed based on a cloud language model. According to the differences between application scenarios and hardware carriers, it can be divided into the following three mainstream technical paths:

[0003] (1) Smart home speaker. This type of device takes the strong correlation architecture of “instruction-response” as the core, and realizes precise control of smart home devices through a preset voice instruction set. It is especially good at handling switching and adjusting, environmental parameter setting and other operation requirements with clear physical mapping. Its technical advantage lies in the control efficiency in structured scenarios, but it depends on fixed power supply environment and home space deployment, and the interaction scene is significantly limited.

[0004] (2) Mobile terminal voice assistant. Relying on the hardware basis of mobile devices such as smart phones, it has portability and multi-modal interaction capability, and can complete transactional operations such as instant inquiry and schedule management, and realize personalized service recommendation through machine learning. It provides seamless experience of “dialogue as a service”, but it is highly dependent on high-performance hardware support of the terminal.

[0005] (3) Cloud service voice assistant. It is mainly applied to professional fields such as intelligent customer service, financial consultation and medical diagnosis, and through the construction of a field-specific language model, it realizes the accurate transmission of professional information while ensuring the smoothness of the dialogue. Its technical characteristics are strong model professionalism, but it needs to rely on fixed hardware carriers for operation, and the cost of single-point deployment is high.

[0006] Although the above technical paths have realized mature application in their respective fields, there are still the following problems when landing voice AI interaction in low-power terminals:

[0007] (1) Cloud centralized processing requires the terminal to maintain network connection continuously, and the communication power consumption accounts for a high proportion, which exceeds the battery carrying capacity of the low-power terminal.

[0008] (2) The local deployment scheme is constrained by the computing power of embedded devices and cannot support real-time inference of complex AI models, resulting in simplified interaction functions.

[0009] (3) In the traditional fixed state mode, the voice collection and processing modules are in the active state all the time, and the invalid energy consumption is significantly increased, further exacerbating the imbalance between endurance and experience.

[0010] In summary, under the core demand of voice AI interaction in low-power terminals, the existing technology cannot simultaneously meet the multiple demands of long endurance, high response and rich content, and a new technical solution is needed to realize functional integrity under energy consumption constraints. SUMMARY

[0011] In view of the above defects of the prior art, the present application proposes a voice AI interaction state management method for low-power terminals, comprising the following steps:

[0012] S1: system initialization and state machine deployment, activate the power management module to stabilize power supply, the communication module reduces the radio frequency power after completing network registration, the audio chip loads low-power configuration, and the state machine of the preset "sleep-standby-wakeup-interaction" four-level state parameters and migration conditions is deployed, and after initialization, the system enters standby state;

[0013] S2: standby state low-power operation, the communication module sends small volume heartbeat packet at low frequency to maintain connection, the audio chip intermittently listens to the wake-up word and cuts off the microphone power supply in the gap, the power management module monitors power consumption, and if there is no wake-up after timeout, it is migrated to sleep state;

[0014] S3: wake-up state activation and intent confirmation, after detecting the key or keyword trigger, the audio chip switches to the recording mode, the voice activity detection module detects the voice activity to confirm the interaction intent, the communication module establishes a high-speed data channel, and if there is no valid voice, it falls back to standby state;

[0015] S4: AI voice interaction process execution, the audio chip compresses and encodes the cached voice, uploads it to the cloud by the communication module, the cloud large model analyzes the semantics and generates a reply fragment, and the terminal activates the decoding and playback after receiving, and waits for subsequent input and falls back to standby state after playback;

[0016] S5: external device control instruction processing, the cloud analyzes the control intent and synchronously pushes the structured instruction, the terminal verifies and sends the control signal to the bound device, the device returns the execution result to the cloud, and the cloud integrates it into a voice reply and issues it;

[0017] S6: state recovery and sleep trigger, after the interaction ends, each module resumes the low-power mode, the system returns to standby state, and if there is no wake-up after standby timeout, the communication module sends a sleep notification and disconnects, cuts off the power supply of non-core modules and enters sleep, and the key trigger restarts the initialization.

[0018] Preferably, the migration conditions include: the sleep-to-standby trigger source is a physical key interrupt; the standby-to-wakeup condition is keyword matching or key trigger; the wakeup-to-interaction basis is that the VAD (voice activity detection) module detects valid voice; the interaction-to-standby threshold is a preset time length without voice input; and the standby-to-sleep threshold is a preset time length without wake-up operation.

[0019] Preferably, the wake-up state activation and intention confirmation include: the audio chip is immediately started for real-time recording when the key trigger; the voice activity detection module detects the subsequent audio for activity detection, and the wake-up word detection channel is closed after confirming the intention.

[0020] Preferably, the AI voice interaction includes a local wake-up and cloud reasoning architecture, audio data is transmitted by fragmentation, and local cache preprocessing is used to realize low-latency interaction; the local wake-up includes triggering and confirming by a hardware-level wake-up word detection and voice activity detection module; the cloud reasoning includes a cloud pre-trained language large model responsible for semantic analysis and content generation.

[0021] Preferably, a dynamic power consumption adjustment algorithm is further used for control in each state; the dynamic power consumption adjustment algorithm includes: in the standby state, the radio frequency module is closed after the communication module completes the heartbeat communication, and the audio chip is powered off in the gap of collection; in the interactive state, the communication module transmission link is closed after uploading data, and the receiving channel is closed when playing voice; in the hibernation state, only the key interrupt detection circuit is reserved.

[0022] Preferably, the method uses a module on-demand start-stop strategy; the on-demand start-stop strategy includes: cutting off the power supply of redundant modules in the inactive stage, and only powering the modules participating in the process in the interactive process.

[0023] Preferably, the structured instruction contains device identification, operation type and parameter information, the terminal checks the validity of the instruction through the AI voice device control module, controls the external device to execute the operation, and feeds back the result data to the cloud through the communication module.

[0024] Preferably, the step S4 further includes: the audio chip converts the collected voice into a 16kHz sampling rate digital audio stream, and the data packet uploaded by the communication module conforms to the cloud transmission protocol.

[0025] The application has the following beneficial effects:

[0026] (1) Through the closed-loop management of the "hibernation-standby-wake-up-interactive" four-level state machine, combined with the dynamic power consumption adjustment algorithm, the intelligent energy efficiency switching in different scenes is realized. The overall power consumption is controlled within 5mA in the standby state, and the current can be as low as 1-3μA in the hibernation state, which prolongs the standby time and improves the continuous running time compared with the traditional continuous wake-up scheme.

[0027] (2) The power supply of redundant modules is accurately cut off in the inactive stage, the communication link uses the on-demand activation strategy, and the LTE module only establishes a high-speed connection after wake-up, which greatly reduces the invalid power consumption and solves the technical contradiction between continuous AI interaction and long endurance.

[0028] (3) Adopting a "local wake-up-cloud reasoning" hybrid architecture, the AI language large model is deployed in the cloud, and the interactive content is defined by the user, breaking through the limitations of traditional fixed instruction sets, covering transaction queries, service consultation, device control and other multi-scene scenarios, solving the problem of lack of interactive content and fixed dimension in existing solutions.

[0029] (4) Hardware-level wake-up word detection and VAD (Voice Activity Detection) module work together to avoid false triggering during standby and quickly complete voice activity judgment after waking up. The cloud uses a pre-trained language large model to analyze semantics, accurately identifies instructions, and simultaneously realizes end-to-end low-latency interaction through audio fragmentation transmission, local cache preprocessing and other methods, ensuring high continuity of user and AI question answering.

[0030] (5) No need to carry high-performance local AI chips, relying on cloud computing power to support complex model reasoning, the core hardware only needs low-cost components such as LTE modules, codec audio chips, basic power management modules, etc., reducing hardware deployment costs and adapting to resource-constrained devices such as wearable health monitoring terminals and Internet of Things nodes.

[0031] The voice AI interaction state management method for low-power terminals of the present application has the characteristics of low cost, strong function and rapid feedback. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The low-power terminal system block diagram of the present application embodiment is shown in the figure;

[0033] Figure 2 The interactive flowchart of the present application embodiment is shown in the figure;

[0034] Figure 3 The state transition diagram of the present application embodiment is shown in the figure. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0036] Embodiment one

[0037] This embodiment takes a low-power terminal as an application carrier, as shown in Figure 1 , Figure 2 The terminal is equipped with an LTE module, a codec audio chip, a microphone, a speaker, a power management chip and a 300mAh lithium battery, and the voice AI interaction and low-power coordination are realized based on the method of the present application, the specific steps are as follows:

[0038] S1: System initialization and state machine deployment.

[0039] After the system is powered on, the initialization process is started. First, the power management chip is activated, the output voltage of the lithium battery is stabilized, and stable power supply is provided for each module. After the LTE module completes network registration and establishes a basic connection with the cloud AI platform, the radio frequency power is immediately reduced to maintain only the minimum link activity. The codec audio chip loads low-power configuration parameters, closes redundant sampling channels, and only retains the basic audio acquisition function required for wake-up word detection.

[0040] At the same time, the state machine control module is deployed, as shown in Figure 3 The four-level state parameters of "sleep-standby-wakeup-interaction" are preset, and the migration conditions and trigger thresholds between states are defined. The trigger source from sleep to standby is physical key interrupt; the trigger condition from standby to wakeup is keyword matching or key trigger; the switching basis from wakeup to interaction is that the VAD (Voice Activity Detection) module detects valid user voice; the fallback threshold from interaction to standby is 10 seconds of no voice input; the timeout threshold from standby to sleep is 30 minutes of no wakeup operation. After initialization is completed, the system automatically enters the standby state, and the overall power consumption is controlled within 5mA.

[0041] S2: Standby state low-power running.

[0042] In the standby state, the system enters the core low-power running phase. The LTE module sends a heartbeat data packet to the cloud every 60 seconds, the data packet size is compressed to within 128 bytes, and the radio frequency module is immediately turned off after communication to avoid energy loss caused by continuous network connection. The codec audio chip is in intermittent listening mode, and the microphone is activated every 200 milliseconds to collect audio segments. The wake-up word is detected by a hardware-level feature matching algorithm, and the microphone power supply is completely cut off during the collection interval.

[0043] The power management chip monitors the current consumption of each module in real time, and continuously samples the power consumption data of the LTE module and the codec chip. If no key operation or wake-up word matching signal is detected for 30 consecutive minutes, the system triggers the state migration instruction, gradually turns off the power supply of non-core modules, and smoothly transitions to the sleep state.

[0044] S3: Wakeup state activation and intent confirmation.

[0045] When a valid trigger signal is detected in the standby state, the system enters the wakeup state. If the trigger source is a physical key, the codec audio chip immediately switches to real-time recording mode and starts continuous audio collection with the microphone; if the trigger source is keyword matching, the VAD module is started synchronously to detect the activity of the audio signal within the next 1.5 seconds to determine whether there is a real interaction intent.

[0046] At this time, the LTE module increases the radio frequency power, and establishes a high-speed data channel with the cloud platform, but does not upload complete audio data. The codec chip converts the collected voice signal into a 16kHz sampling rate digital audio stream, temporarily stores it through the internal buffer, and at the same time, closes the wake-up word detection channel to avoid waste of computing power caused by repeated triggering. If the VAD module detects no valid voice input, the system automatically falls back to the standby state after 10 seconds.

[0047] S4: AI voice interaction process execution.

[0048] After detecting the user voice in the wake-up state, the system enters the interaction state. The codec chip compresses and encodes the buffered audio stream to generate data packets conforming to the cloud transmission protocol, which are uploaded to the cloud AI platform by the LTE module. After uploading is completed, the LTE module closes the data transmission link and only keeps the receiving channel to wait for cloud feedback, and the power management chip cuts off the power supply of the sensor modules not involved in the current process.

[0049] After receiving the data packet, the cloud AI platform performs semantic analysis and intent recognition through the pre-trained language large model, generates voice reply content and converts it into audio data, and issues it to the terminal in the form of fragments. After receiving the fragmented data, the LTE module triggers an interrupt, activates the audio decoding function of the codec chip, and the loudspeaker plays the reply voice. During the playing period, the system closes the receiving channel of the LTE module, and only maintains the power supply of the codec chip and the loudspeaker, further reducing the energy consumption in the interaction process.

[0050] After the voice playback is completed, the codec chip keeps the decoding module active for 3 seconds to wait for possible subsequent questions from the user. If no new voice input is detected, it immediately switches to a low-power listening mode, the LTE module re-establishes the heartbeat connection, and the system falls back to the standby state.

[0051] S5: External device control instruction processing.

[0052] During the interaction process, if the cloud AI platform analyzes the user's device control intent, it will simultaneously push the device control instruction when issuing the voice reply data. The instruction is transmitted in the form of structured data, including device identification, operation type and parameter information, for example, the instruction data corresponding to the user's instruction "turn the body LED lamp into warm yellow" contains the LED device identification, the on state and the color parameters of red 180, green 120 and blue 50.

[0053] After receiving the instruction, the LTE module triggers the AI voice device control module to check the instruction, and sends a control signal to the LED lamp after confirming that the LED lamp is a valid device. After the external device performs the operation, the result feedback data is generated and returned to the cloud platform through the LTE module. The cloud platform integrates the operation result into voice reply content and sends it to the terminal through subsequent audio data packets, realizing the closed-loop control of "instruction-execution-feedback".

[0054] S6: state recycling and hibernation triggering.

[0055] After the interaction process is completed, the system starts the state recycling mechanism. The codec chip clears the audio buffer and returns to the low-power listening mode; the LTE module disconnects the high-speed data connection and returns to the heartbeat communication mode; the power management chip re-provides power for the previously closed sensor modules, and the terminal returns to the standby state.

[0056] If no wake-up trigger signal is detected for 30 consecutive minutes in the standby state, the system starts the hibernation process. First, the LTE module sends a hibernation notification to the cloud, disconnects the network connection and turns off the radio frequency module. The power management chip successively cuts off the power supply of the LTE module, the codec audio chip and the microphone, and only the interrupt detection circuit of the physical button is retained. At this time, the terminal enters a deep hibernation state, and the actual current is stable at about 2 mu A, which can realize a standby endurance of up to 30 days. When the user presses the physical button, the power management chip reactivates the power supply of each module, and the system repeats the initialization process and enters the standby state again.

[0057] The above only describes the preferred embodiments of the present application, and it should be noted that for ordinary skilled persons in the technical field, some improvements and refinements can be made without departing from the technical principles of the present application, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1.A voice AI interaction state management method for a low-power terminal, the method comprising: Comprise the following steps: S1: system initialization and state machine deployment, activate power management module voltage stabilization power supply, communication module completes network registration and reduces radio frequency power, audio chip loads low power consumption configuration, deployment of preset "sleep- standby- wake up- interaction" four level state parameters and migration conditions state machine, after initialization system enters standby state; S2: standby state low power consumption operation, communication module sends small volume heartbeat packet at low frequency to maintain connection, audio chip intermittently listens to wake-up word and cuts off microphone power supply in gap, power management module monitors power consumption, and migrates to sleep state if there is no wake-up after timeout; S3: wake-up state activation and intent confirmation, after detecting key or keyword trigger, audio chip switches to recording mode, voice activity detection module detects voice activity to confirm interaction intent, communication module establishes high-speed data channel, and falls back to standby state if there is no valid voice; S4: AI voice interaction process execution, audio chip compresses and encodes cached voice, uploads to cloud end by communication module, cloud end large model analyzes semantics and generates reply fragments, terminal activates decoding and playing after receiving, and waits for subsequent input and falls back to standby state after playing; S5: external device control instruction processing, cloud end analyzes control intent and synchronously pushes structured instruction, terminal verifies and sends control signal to bound device, device returns execution result to cloud end, cloud end integrates into voice reply and issues; S6: state recycling and sleep triggering, after interaction ends, each module restores low power consumption mode, system returns to standby state, communication module sends sleep notification and disconnects if there is no wake-up after standby timeout, cuts off power supply of non-core modules and enters sleep state, and key trigger restarts initialization. 2.The voice AI interaction state management method for a low-power terminal of claim 1, wherein, The migration conditions comprise: sleep to standby trigger source is physical key interruption; standby to wake-up condition is keyword matching or key trigger; wake-up to interaction is based on voice activity detection module detecting valid voice; interaction to standby threshold is preset time length without voice input; standby to sleep threshold is preset time length without wake-up operation. 3.The voice AI interaction state management method for a low-power terminal of claim 1, wherein, The wake-up state activation and intent confirmation comprise: audio chip starts real-time recording immediately when key is triggered; voice activity detection module detects subsequent audio activity when keyword is triggered, and closes wake-up word detection channel after confirming intent. 4.The voice AI interaction state management method for a low-power terminal of claim 1, wherein, The AI voice interaction comprises: local wake-up and cloud reasoning architecture, audio data is transmitted by fragmentation, local cache preprocessing realizes low latency interaction; the local wake-up comprises local trigger confirmation by hardware level wake-up word detection and voice activity detection module; the cloud reasoning comprises cloud pre-training language large model responsible for semantic analysis and content generation. 5.The voice AI interaction state management method for a low-power terminal of claim 1, wherein, Further comprise dynamic power consumption adjustment algorithm for control in each state; the dynamic power consumption adjustment algorithm comprises: in standby state, communication module closes radio frequency module after completing heartbeat communication, and audio chip is powered off during gap; In sleep state, only key interruption detection circuit is reserved. 6.The voice AI interaction state management method for a low-power terminal of claim 1, wherein, The method adopts module on-demand start-stop strategy; the on-demand start-stop strategy comprises: cutting off power supply of redundant modules in non-active stage, and only supplying power to modules participating in process in interaction process. 7.The voice AI interaction state management method for a low-power terminal of claim 1, wherein, The structured instruction includes device identification, operation type and parameter information, the terminal checks the instruction validity through the AI voice device control module, controls the external device to execute the operation, and feeds back the result of the data to the cloud through the communication module. 8.The voice AI interaction state management method for a low-power terminal of claim 1, wherein, The step S4 further includes that the audio chip converts the collected voice into a digital audio stream with a sampling rate of 16 kHz, and the data packet uploaded by the communication module conforms to the cloud transmission protocol.

Images